incubator-kato-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Lukasz <flo...@intercel.com.pl>
Subject Re: Processing huge heap dump.
Date Fri, 15 Jan 2010 18:18:35 GMT
Hi Steve,

Actually I've made two changes.

First was increasing the size of array in classes BitMaskMappingArray 
and ArrayBitMaskMappingStrategy, which is probably the same what you 
have made.
I had similar idea about using heuristic/prediction of how many records 
can be in a dump.
Other (or additional) solution could be increasing cacheSize instead of 
bucketSize when cache will be exhausted.

I have filling that main purpose of those classes is providing mapping 
between "index of record" and "absolute position of the record in the 
file".
And if my filling are correct, the mapping comprise of:
1) finding absolute location in the file
2) skipping a number of records (called "seek") before we set reader's 
position just on the beginning of required record.
And the second step is the problem (I believe), because the bigger 
bucket we have, the more records we have to skip (methods called: 
"moveRelativeElement" are responsible).
I think this, and the fact that bucketSize is doubled when cache is 
exhausted, are the reasons why performance drops over the time.

To fix it I changed following parts of code in 
org.apache.kato.hprof.datalayer.HProfFile:
        public HeapDumpHProfRecord(short tag, IDataProvider dataProvider)
                throws IOException {
            ...
            subRecordArray = new BitMaskMappingArray(50000000, 25,
                    new HProfDumpRecordArrayProvider(dataProvider));
        }
and
    private void buildHeader() throws IOException {
        ...
        array = new BitMaskMappingArray(1000000, 25, new 
HPROFRecordProvider(
                subset));

    }


Second problem was incorrect size of HeapDumpHProfRecord which was read 
from the file (recordSize overflow problem).
I didn't find general solution for this problem. 
I just assumed that HeapDumpHProfRecord will end with the end of the 
file, which is true in my case.

To achieve this I made following changes (also in HProfFile):
                case 0x0c:
                    long l = recordProvider.dataRemaining();
                    IHeapDumpHProfRecord rec = new HeapDumpHProfRecord(tag,
                            new SubsetDataProvider(recordProvider, l));
                    recordProvider.moveBy(l);
                    return rec;

I also had to change signature of "moveBy" method to accept long instead 
of int.

I'll try to send you part of my source code together with .svn metadata, 
so you should be able to find what exactly I've changed.

Regards
Lukasz



Steve Poole wrote:
> I've run some performance tests on various configs for HProfFile caching.
> See here
> http://spreadsheets.google.com/pub?key=tes0VC-Wyz-lTe9sKNen2Hw&output=html
>
> Our current setting for cache size and bucket size is 1000,25  - Bucket size
> does effect the speed of analysis but its cache size that really does it!
> We need to configure the cache size based on some heuristic about the size
> of the dump file.
>
> On Thu, Jan 14, 2010 at 5:21 PM, Lukasz <flower@intercel.com.pl> wrote:
>
>   
>> Hi Steve, Stuart,
>>
>> In last two days I was trying to move further.
>>
>> I tried to prepare some workaround for incorrect length.
>> First idea was to utilize the fact that integer overflow probably took
>> place, so I tried to search for 0x2c just at overflow points (probably quite
>> brittle approach). But it turned out, that in my heap dumps (both real one
>> and prepared for development purposes) there is no 0x2c byte after/at the
>> end of the dump.
>> I've noticed however that for both dumps the heap record ends at the end of
>> the file. Probably it is not general rule, but for my purposes it should be
>> enough, so I changed kato sources slightly to utilize this and decided to
>> move further.
>>
>> So far I have some code which extracts static data from dump (strings,
>> classes) and searches for instances of given class.
>> I have run it against real dump. It took around 40 min (to get static data)
>> and 50 min (to find instances), which is not bad I think.
>> I didn't made careful reasoning about performance but CPU (not I/O) looked
>> to be a bottleneck. Performance also looked to be quite stable across whole
>> dump.
>>
>> 60GB dump was extracted for core file using jmap (I don't remember exactly
>> from which version of java this jmap came from, but for sure it was Java
>> 1.6).
>> Following JVM version was dumped to core file:
>> java version "1.6.0_10"
>> Java(TM) SE Runtime Environment (build 1.6.0_10-b33)
>> Java HotSpot(TM) 64-Bit Server VM (build 11.0-b15, mixed mode)
>>
>> I was also suspecting that heap dump segments were introduced to handle
>> +4GB heap sizes, but it looks like sometimes it can not be a case.
>>
>> Anyway, it is nice that I was able to go through whole dump without any
>> error.
>>
>> Regards
>> Lukasz
>>
>>
>> Steve Poole wrote:
>>
>>     
>>>  Hi Lukasz,
>>>
>>>  You say your  hprof format is 1.0.1   so , according to the information I
>>> have about hprof format,  that means your won't have any heap dump
>>> segments
>>> in your file.  Apparently they are in the  1.0.2 format which is a Java 6
>>> addition and is only used when the heap is too big for the old format to
>>> cope with.
>>>
>>> What JVM and OS are you using to produce this dump?
>>>
>>> On Tue, Jan 12, 2010 at 6:17 PM, Stuart Monteith <stukato@stoo.me.uk>
>>> wrote:
>>>
>>>
>>>
>>>       
>>>> Hi,
>>>>
>>>>
>>>> You are quite right, 4 bytes for the record length. If we had the file
>>>> format wrong for 64 bit
>>>> hprof dumps we'd not be able to read any dump, so it must be an overflow
>>>> of
>>>> the record length.
>>>> In that case, the only possible solution I can think of would be for us
>>>> to
>>>> change to code to search
>>>> for the HEAP DUMP END record (tagged 0x2c) to detect the end of the HEAP
>>>> DUMP.
>>>>
>>>> Of course, there was some puzzling over the difference between HEAP DUMP
>>>> SEGMENT (0x1c) and HEAP DUMP
>>>> (0x0c) before.
>>>>
>>>> I'd expect we'd get:
>>>>
>>>>   HEAP DUMP
>>>>       heap dump records....
>>>>   HEAP DUMP END
>>>>
>>>> or
>>>>
>>>>   HEAP DUMP SEGMENT
>>>>       heap dump records
>>>>   HEAP DUMP SEGMENT
>>>>       heap dump records
>>>>      :      :
>>>>   HEAP DUMP END
>>>>
>>>>
>>>> If we ignore the record size for the HEAP DUMP and calculate it
>>>> ourselves,
>>>> we might have more luck.
>>>>
>>>> Regards,
>>>>   Stuart
>>>>
>>>>
>>>>
>>>> Lukasz wrote:
>>>>
>>>>
>>>>
>>>>         
>>>>> Hello,
>>>>>
>>>>> Since performance looks not bad after changing array size, today I have
>>>>> tried to process my 60GB heap dump. Unfortunately, very fast I've ended
>>>>> with
>>>>> exception:
>>>>> Exception in thread "main" java.lang.IllegalArgumentException: range
>>>>> -1293366370 is less than 0
>>>>> at
>>>>>
>>>>> org.apache.kato.common.SubsetDataProvider.<init>(SubsetDataProvider.java:29)
>>>>>
>>>>> at
>>>>>
>>>>> org.apache.kato.hprof.datalayer.HProfFile$HPROFRecordProvider.getCurrentElement(HProfFile.java:487)
>>>>>
>>>>> at
>>>>>
>>>>> org.apache.kato.hprof.datalayer.HProfFile$HPROFRecordProvider.getCurrentElement(HProfFile.java:437)
>>>>>
>>>>> at
>>>>>
>>>>> org.apache.kato.common.BitMaskMappingArray.get(BitMaskMappingArray.java:56)
>>>>> at
>>>>> org.apache.kato.hprof.datalayer.HProfFile.getRecord(HProfFile.java:784)
>>>>> ... (my code) ...
>>>>>
>>>>> It looks like range of HeapDumpHProfRecord had value: -1293366370.
>>>>>
>>>>> Following is the part of the dump which I believes is responsible for
>>>>> that
>>>>> (B2 E8 CB 9E):
>>>>> 0055DD84 25 00 00 2A │ B7 AE 22 06 │ 80 00 00 00 │ 01 00 00 2A
│ B7 AE
>>>>> 20
>>>>> 0B │ 88 02 00 00 │ 00 00 00 00 │ 00 18 00 00 │ 23 26 00 00
>>>>> 0055DDA8 2A B7 B0 A1 │ 64 28 00 00 │ 00 01 00 00 │ 2A B7 B0 33
│ EA F8
>>>>> 0C
>>>>> 00 │ 00 00 00 B2 │ E8 CB 9E 20 │ 00 00 2A B7 │ B0 DE AF C8
>>>>> 0055DDCC 00 00 00 01 │ 00 00 2A B7 │ AE 22 0F 20 │ 00 00 2A AA
│ BE 21
>>>>> 3D
>>>>> 28 │ 00 00 00 00 │ 00 00 00 00 │ 00 00 2A AA │ BE 25 B7 D8
>>>>>
>>>>> What is interesting hprof file format reserves only 4 bytes for length
>>>>> of
>>>>> a record, which allows to express record whose size is at most 4GB
>>>>> (assuming
>>>>> unsigned 4 bytes).
>>>>>
>>>>> My dump's version is: JAVA PROFILE 1.0.1
>>>>>
>>>>> Regards
>>>>> Lukasz
>>>>>
>>>>>
>>>>> Lukasz wrote:
>>>>>
>>>>>
>>>>>
>>>>>           
>>>>>> Hi Stuart, Steve,
>>>>>>
>>>>>> I've taken deeper look into code. I still didn't trace carefully
index
>>>>>> calculation in classes BitMaskMappingArray and
>>>>>> ArrayBitMaskMappingStrategy,
>>>>>> but I managed to improve performance by increasing arrays size in
those
>>>>>> classes (which is set in HProfFile class).
>>>>>>
>>>>>> If I understand code correctly, when capacity of BitMaskMappingArray
>>>>>> will
>>>>>> be exhausted bucketSize is doubled, which in turn causes that more
>>>>>> reads
>>>>>> (even cached) is required to set position of
>>>>>> IDataProvider/IArrayEntryProvider.
>>>>>>
>>>>>> Following are loading time results for default array size (1000)
and
>>>>>> increased (1000000). Test ran against generated dump file (5000000
>>>>>> instances
>>>>>> of Data).
>>>>>> Default (1000):
>>>>>> HeapSubRecord: 100000 (866ms, 4215kB)
>>>>>> HeapSubRecord: 200000 (1716ms, 7879kB)
>>>>>> HeapSubRecord: 300000 (2833ms, 11263kB)
>>>>>> HeapSubRecord: 400000 (3889ms, 14283kB)
>>>>>> HeapSubRecord: 500000 (3893ms, 17319kB)
>>>>>> HeapSubRecord: 600000 (7248ms, 20479kB) (here probably buckedSize
was
>>>>>> doubled)
>>>>>> HeapSubRecord: 700000 (7721ms, 23531kB)
>>>>>> HeapSubRecord: 800000 (7729ms, 26567kB)
>>>>>> HeapSubRecord: 900000 (7731ms, 29671kB)
>>>>>> HeapSubRecord: 1000000 (7704ms, 32731kB)
>>>>>> ... (I didn't wait until end)
>>>>>>
>>>>>> Increased(1000000):
>>>>>> HeapSubRecord: 100000 (622ms, 17809kB)
>>>>>> HeapSubRecord: 200000 (309ms, 20345kB)
>>>>>> HeapSubRecord: 300000 (283ms, 23861kB)
>>>>>> HeapSubRecord: 400000 (274ms, 27921kB)
>>>>>> HeapSubRecord: 500000 (269ms, 29957kB)
>>>>>> HeapSubRecord: 600000 (264ms, 31993kB)
>>>>>> HeapSubRecord: 700000 (272ms, 36097kB)
>>>>>> HeapSubRecord: 800000 (288ms, 37739kB)
>>>>>> HeapSubRecord: 900000 (263ms, 39835kB)
>>>>>> HeapSubRecord: 1000000 (259ms, 41931kB)
>>>>>> HeapSubRecord: 1100000 (300ms, 44773kB)
>>>>>> HeapSubRecord: 1200000 (283ms, 46901kB)
>>>>>> HeapSubRecord: 1300000 (291ms, 49029kB)
>>>>>> HeapSubRecord: 1400000 (328ms, 53801kB)
>>>>>> HeapSubRecord: 1500000 (259ms, 53801kB)
>>>>>> HeapSubRecord: 1600000 (272ms, 58125kB)
>>>>>> HeapSubRecord: 1700000 (264ms, 60293kB)
>>>>>> HeapSubRecord: 1800000 (264ms, 62473kB)
>>>>>> HeapSubRecord: 1900000 (361ms, 61373kB)
>>>>>> HeapSubRecord: 2000000 (274ms, 63105kB)
>>>>>> ...
>>>>>> HeapSubRecord: 9000000 (284ms, 231969kB)
>>>>>> HeapSubRecord: 9100000 (272ms, 233597kB)
>>>>>> HeapSubRecord: 9200000 (281ms, 236357kB)
>>>>>> HeapSubRecord: 9300000 (274ms, 240469kB)
>>>>>> HeapSubRecord: 9400000 (279ms, 244541kB)
>>>>>> HeapSubRecord: 9500000 (269ms, 246549kB)
>>>>>> HeapSubRecord: 9600000 (279ms, 250565kB)
>>>>>> HeapSubRecord: 9700000 (265ms, 252573kB)
>>>>>> HeapSubRecord: 9800000 (279ms, 256629kB)
>>>>>> HeapSubRecord: 9900000 (265ms, 258669kB)
>>>>>> HeapSubRecord: 10000000 (463ms, 263997kB)
>>>>>> (end)
>>>>>>
>>>>>> i.e. my 60GB dump file contains more than 1 100 000 000 of objects
(if
>>>>>> I
>>>>>> remember correctly).
>>>>>>
>>>>>> Regards
>>>>>> Lukasz
>>>>>>
>>>>>>
>>>>>> Stuart Monteith wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>             
>>>>>>> The hprof dump reader spends a lot of time reading the whole
file, for
>>>>>>> various reason.
>>>>>>> The indices it has in memory are constructed through an initial
read,
>>>>>>> and this is also
>>>>>>> the source of the memory usage. In addition, there is some correlation
>>>>>>> to be done which
>>>>>>> also takes up time, and induces yet more reading.
>>>>>>>
>>>>>>> I'm sure some work could be done to improve the performance further,
>>>>>>> but
>>>>>>> we'll have to
>>>>>>> look at the tradeoff between diskspace and memory usage. The
hprof
>>>>>>> file
>>>>>>> format itself
>>>>>>> is what it is, however, and we have no influence over that. The
CJVMTI
>>>>>>> agent is has lots of
>>>>>>> room for improvement, but I suspect its potential for improvement
is
>>>>>>> unlikely to be much better
>>>>>>> than existing hprof implementations. The built-in JVM hprof dumper
>>>>>>> will
>>>>>>> probably be a hard act
>>>>>>> to follow.
>>>>>>>
>>>>>>> The HProf implementation is not thread-safe. Realistically, I
think it
>>>>>>> is something that ought to
>>>>>>> be considered once things are more mature. There will be algorithms
>>>>>>> that
>>>>>>> can deal with the JVM
>>>>>>> structure sensible.
>>>>>>>
>>>>>>> And thanks Lukasz, it's great to have your input.
>>>>>>>
>>>>>>> Regards,
>>>>>>> Stuart
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>               
>>>>>>             
>>>>> --
>>>>>           
>>>> Stuart Monteith
>>>> http://blog.stoo.me.uk/
>>>>
>>>>
>>>>
>>>>
>>>>         
>>>
>>>
>>>       
>>     
>
>
>   


Mime
View raw message